Artificial intelligence as a formal field was founded in 1956 at the Dartmouth workshop. While researchers like Alan Turing explored machine intelligence in the 1950s, the term "artificial intelligence" was officially coined by John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester in a 1955 proposal.
The Birth of the Field (1956)
The origins of AI are not tied to a single inventor, but rather to a pivotal moment in academic history. In 1955, a group of researchers—John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester—wrote a proposal that introduced the term "artificial intelligence" to the world.
This proposal led to the 1956 Dartmouth workshop, which is widely considered the founding event of AI as a formal academic discipline. Before this official naming, the groundwork was being laid by thinkers like Alan Turing. In 1950, Turing published "Computing Machinery and Intelligence," where he proposed a method for testing machine intelligence known as the "imitation game," now famously called the Turing test.
The Era of Logic and Rules
From the 1950s through the 1970s, the field was dominated by what is known as "Symbolic AI." During this period, researchers operated under the belief that intelligence could be captured by writing down explicit logic and rules for a machine to follow.
This era saw several important milestones in how humans interacted with machines:
- Logic Theorist: One of the earliest successful programs designed to mimic human reasoning.
- ELIZA: Created by Joseph Weizenbaum at MIT in 1966, this was a pattern-matching chatbot. It simulated a psychotherapist, and its ability to mimic conversation surprised many users of the time.
However, this era of high expectations eventually hit a wall. In the mid-1970s, the field experienced its first "AI winter." This was a period where funding was significantly cut because the early promises of AI researchers outran the actual technical results available at the time. For more on these cycles, see our overview of the AI winter.
Expert Systems and the Second Winter
By the 1980s, the approach to AI shifted toward "Expert Systems." Instead of trying to solve general intelligence, developers focused on narrow, specific domains. These systems, such as XCON, were designed to encode the specialized knowledge of human experts into a set of rules to solve industrial problems, such as configuring complex computer orders.
While these systems showed that AI could save real money in controlled business environments, they had a major flaw: they were "brittle." Because they relied on hand-written rules, they worked well in stable environments but broke down when faced with the messy, unpredictable nature of the real world. This limitation led to a second "AI winter" in the late 1980s and early 1990s, as the market for specialized AI hardware collapsed and the cost of maintaining these rule-based systems became too high.
| Era | Core Approach | Key Characteristic |
|---|---|---|
| 1950s–1970s | Symbolic AI | Focus on logic and search over symbols. |
| 1980s | Expert Systems | Encoding specialist rules for narrow domains. |
The Shift to Machine Learning
Between the 1990s and the 2000s, the philosophy of AI underwent a fundamental change. Instead of humans trying to write down every rule for a computer to follow, researchers began focusing on machine learning. In this paradigm, computers learn to recognize patterns directly from data.
A landmark moment in this era occurred in 1997, when IBM's Deep Blue defeated the world chess champion, Garry Kasparov. Interestingly, Deep Blue achieved this victory primarily through massive search capabilities rather than true "learning" in the modern sense. Despite this, the era marked a clear move away from human-coded logic and toward statistical pattern recognition.
The Deep Learning and GPU Revolution
The landscape of AI changed again around 2012. This was not caused by a single discovery, but by the convergence of three critical factors:
- Large Datasets: The availability of massive amounts of labeled data, such as the ImageNet dataset.
- Better Training Methods: New ways to train multi-layer neural networks.
- Cheap Compute: The rise of powerful GPUs (Graphics Processing Units) originally designed for gaming, which were found to be perfect for the heavy math required by neural networks.
The success of AlexNet in the 2012 ImageNet challenge proved that deep convolutional networks were the new gold standard. This era saw the rise of neural networks capable of recognizing complex images and even winning high-level games, such as DeepMind's AlphaGo, which defeated Lee Sedol in 2016.
Transformers and Generative AI
In 2017, Google researchers published a paper titled "Attention Is All You Need," which introduced the "Transformer" architecture. This was a breakthrough because it allowed models to process entire sequences of data in parallel, rather than one piece at a time. This architecture became the foundation for what we now call "foundation models."
This led to a rapid evolution in how we interact with AI:
- Pre-training: Models like GPT-3 (released in 2020) showed that if you train a model on a massive amount of text, it can perform new tasks just by being given a few examples in a prompt.
- Generative AI: In 2022, the release of ChatGPT brought conversational AI to the general public, making generative models accessible to everyone.
For a deeper look at how we reached this point, you can explore the evolution of AI.
The Move Toward AI Agents
As we move through 2024 and 2025, the focus is shifting once again. We are moving from "chatbots"—which are primarily designed for conversation—to "AI agents."
Modern AI agents are designed to do more than just talk; they are built to plan, use software tools, and complete multi-step tasks autonomously. For example, instead of just writing an email, an agent might research a topic, find the correct contact, and execute a workflow across different software tools. This is supported by new standards like the Model Context Protocol (MCP), which helps these agents connect to different data sources and tools more easily.
Questions people ask
Who is considered the father of artificial intelligence?
While many contributed, John McCarthy is credited with coining the term 'artificial intelligence' in 1955 and organizing the seminal 1956 Dartmouth workshop.
What was the Turing Test?
Proposed by Alan Turing in 1950, it is a method of determining a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.
What is an AI winter?
An AI winter refers to periods, such as the mid-1970s and late 1980s, where interest and funding for AI research dropped significantly due to unmet expectations.
What changed AI in 2012?
The breakthrough was the combination of large datasets, new training methods, and the use of GPUs to power deep learning neural networks.
What is the difference between a chatbot and an agent?
Chatbots primarily focus on conversation, whereas agents are designed to use tools, follow multi-step plans, and complete specific tasks autonomously.